stella usage

Totals grouped by org, provider, and model, across every project copied into the hub: calls, tokens, cache reads, cost, and how many different projects contributed. This is the default.

Copies this project's usage data into the hub, picking up where it left off. Safe to run again; --all repairs every project the hub has ever seen.

Keeps the hub from growing forever, by age, by a row limit, or by clearing out projects whose folder is gone.

text prints lined-up columns with a TOTAL row at the bottom (table also works as another name for the same thing). json (wrapped in a versioned envelope) and csv are meant for piping into other tools. The three text columns in CSV follow the RFC-4180 escaping rule, so an org id or model name containing a comma can never shift the columns after it.

Only shows rows recorded under this org id. Rows written before you registered have no org attached, so an --org filter always leaves those out.

Shows tool reliability instead of spend: calls, errors, and the error rate for each (tool, surface) pair, across every project, read from the hub's tool summary. Each row also breaks its errors down by ErrorClass (errors_by_class in JSON/CSV, CLASSES in the table view; unclassified covers calls recorded before their error type was tracked), so a question like "bash's error rate, not counting invalid_input" is simple math, not text matching. This gives you the same numbers behind the Observatory leaderboard, but with no time window and in a form you can pipe to other tools. It cannot be combined with --org, because this summary is keyed by project, not by org, and silently ignoring the filter would give you a wrong answer. It's called --by-tool rather than --tools, because --tools is already used for the session-wide tool-policy setting.

Walks every project the hub already knows about, instead of just the current folder. Use this to repair things: if a project's background sync failed partway through, its position falls behind, and --all fixes every project like that in one pass. A registered project whose local store.db is missing is reported as skipped, not as an error.

Deletes rows older than this window. Use a number followed by d, w, h, mo, or y, for example 90d, 12w, 720h, 3mo, 1y. A bare number means days. The value must be positive; 0 would delete everything up to right now, which is almost always a mistake.

A hard cap on how many rows the hub keeps. The oldest deletable rows are removed until the hub is at or under this number.

Also removes everything the hub holds for a project whose folder is gone and that was never registered to an org: its usage data, its summaries, its saved sync position, and its entry in the hub. "Gone" means the project's own folder is missing while its parent folder still exists, since that's what a real deletion looks like. A whole drive being unavailable, like an unplugged drive or an offline network share, is deliberately not treated as a deletion, so a temporarily disconnected disk never wipes out your local history. A project that comes back gets re-added on its next sync.

Shows what would be deleted, without deleting anything. Worth running first.

Runs SQLite's VACUUM afterward, giving the freed disk space back to your filesystem, and resets the cloud sync positions to match. This also runs automatically after a large prune.

Also deletes rows that haven't been confirmed as sent to the cloud yet, which breaks a pending upload. Off by default.

One project. Reads <workspace>/.stella/private/store.db, the original source of the data, including the per-run cache-cost columns. Its prune only affects that file.

Every project. Reads ~/.stella/usage.db, which is built from the file above. Its prune only affects the hub. Pruning one never touches the other.